Multi-sensor Gait Analysis for Gender Recognition
Abeer Mostafa
1 a
, Toka Ossama Barghash
1
, Asmaa Al-Sayed Assaf
1
and Walid Gomaa
1,2 b
1
Cyber-Physical Systems Lab, Egypt Japan University of Science and Technology, Alexandria, Egypt
2
Faculty of Engineering, Alexandria University, Alexandria, Egypt
Keywords:
Gender Recognition, IMU, Wavelet Transform, Supervised Learning.
Abstract:
Gender recognition has been adopted recently by researchers due to its benefits in many applications such as
recommendation systems and health care. The rise of using smart phones in everyday life made it very easy
to have sensors like accelerometer and gyroscope in phones and other wearable devices. Here, we propose
a robust method for gender recognition based on data from Inertial Measurement Unit (IMU) sensors. We
explore the use of wavelet transform to extract features from the accelerometer and gyroscope signals along
side with proper classifiers. Furthermore, we introduce our own collected dataset (EJUST-GINR-1) which
contains samples from smart watches and IMU sensors placed at eight different parts of the human body.
We investigate which sensor placements on the body best distinguish between males and females during the
activity of walking. The results prove that wavelet transform can be used as a reliable feature extractor for
gender recognition with high accuracy and less computations than other methods. In addition, sensors placed
on the legs and waist perform better in recognizing the gender during walking than other sensors.
1 INTRODUCTION
Gender recognition has been studied widely in the
last decade. Various types of data have been used to
recognise the gender of a person such as images, voice
signals or inertial measurements based on the motion
of the person (Lu et al., 2014), (Garofalo et al., 2019)
and (Zhang et al., 2017). There are many useful ap-
plications that depend on gender recognition such as
speech recognition (Yuchimiuk, 2007), recommenda-
tion systems (Shepstone et al., 2013), and most im-
portantly health care applications (Rosli et al., 2017).
However, there is a huge lack of datasets and accuracy
in the methods that are developed for gender recogni-
tion and the analysis of the data itself. Inertial Mea-
surement Units (IMUs) are known to be embedded in
many wearable devices which lead to useful applica-
tions. It will be convenient to recognise gender based
on their readings (accelerometer, gyroscope, etc).
Datasets collected from IMU sensors are not al-
ways publicly available and most publicly available
datasets don’t focus on diversity of sensor placements
on the human body to get the accelerometer and gyro-
scope signals. For these reasons, we introduce a new
dataset (EJUST-GINR-1) which is collected from col-
a
https://orcid.org/0000-0002-8971-4311
b
https://orcid.org/0000-0002-8518-8908
lege students to record accelerometer and gyroscope
signals from their walking activity. We record signals
from smart watches and IMU sensors placed at eight
different parts of the human body. We study which
part of the human body effectively and uniquely iden-
tifies the gender. We run experiments on each sensor
individually and also on combinations of sensors to
see their effect on the classification accuracy, and in
general we analyse the reliability of each body part in
uniquely determining the gender of the person from
the inertial movements of the corresponding body part
during walking. We run experiments on a different
dataset and analyse the cultural effect that can be im-
portant in changing the nature of the data. Further-
more, we propose a reliable approach to do feature
extraction followed by classification to recognise the
gender based on IMU readings.
There are many approaches to extract relevant fea-
tures used for classification. Recently the most promi-
nent approach is using deep neural networks. How-
ever, these methods perform well when there is a huge
amount of data. This size of data is not always avail-
able when the recognition is based on data coming
from sensors because the process of collecting the
data and annotating it takes much time and effort.
Moreover, the process may require the participation
of many people and the availability of the sensors may
be limited. Accordingly, we propose the use of a fea-
Mostafa, A., Barghash, T., Assaf, A. and Gomaa, W.
Multi-sensor Gait Analysis for Gender Recognition.
DOI: 10.5220/0009792006290636
In Proceedings of the 17th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2020), pages 629-636
ISBN: 978-989-758-442-8
Copyright
c
 2020 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
629

ture extraction method that is both robust and is less
dependable on the amount of available data.
Wavelet transform is a very powerful tool for the
analysis and classification of signals and specifically,
timeseries (Abdu-Aguye and Gomaa, 2019). How-
ever, it is unfortunately not popular within the field of
data science compared with deep learning. Here, we
explore the use of Wavelet transforms for feature ex-
traction along with two classifiers: random forest and
convolutional neural network (CNN) to get a reliable
system for gender recognition.
The rest of the paper is organized as follows. In
section 2, we do literature review of gender recog-
nition. Then, we show some of the work that has
been done using Wavelet transform specially on the
signals coming from IMU sensors. In addition, we
illustrate some of the work where random forest clas-
sifier has proven to be very effective on IMU signals.
In section 3, we introduce our own collected dataset
(EJUST-GINR-1 Dataset). Then, in section 4, we ex-
plain our methodology in detail. In section 5, we
present the setup for our experiments done on our own
collected dataset and the OU-ISIR Gait dataset (Ngo
et al., 2014) respectively. In section 6, we discuss the
main results we accomplished. Finally, we summa-
rize our paper and show potential future work.
2 RELATED WORK
In this section, we review some of the related work
and categorize them into three main categories. Re-
search that focuses on the gender recognition prob-
lem, research techniques that apply Wavelet trans-
form on IMU signals and research that adopts random
forests or CNNs in the classification of timeseries.
2.1 Gender Recognition
Gender recognition has been adopted by researchers
for many years. The variability of sensor types and
applications makes it very wide and difficult research
area.
The authors in (Ngo et al., 2019) presented a com-
petition on gender and age recognition based on sig-
nals of IMU sensors placed on the waist of the per-
son. The evaluation of models was according to per-
formance on the OU-ISIR Gait dataset (Ngo et al.,
2014). They summarize the results of gender recog-
nition of all teams which show that most methods re-
sulted in either a biased or inaccurate results on that
dataset. The best solution used the orientation in-
dependent AE-GDI representation combined with a
CNN which resulted in a classification accuracy up to
75.77%, their solution is presented in (Garofalo et al.,
2019).
In the work produced by (Lu et al., 2014), the
authors proposed a model to do gender recognition
based on computer vision. The model tried to predict
the gender of a person given a sequence of frames
including arbitrary walking directions. They evalu-
ated their model on a dataset consisting of 20 sub-
jects (13 males and 7 females), each was captured in
4 videos. The authors reported the performance of
their model which was promising for computer vision
applications.
A deep learning method was used in (Zhang et al.,
2017) to estimate the age and gender of a person from
face images. The authors used residual networks of
residual networks (RoR) as their model. The model
was pre-trained on ImageNet, then it was fine-tuned
on the IMDB-WIKI-101 data set for learning more
complex features of face images and finally, trans-
fer learning was done on Adience dataset. The RoR
model yielded significant results compared with other
deep learning techniques.
With reference to the work presented by (Jain and
Kanhangad, 2016), the authors investigated solving
the gender recognition problem based on data from
accelerometer and gyroscope sensors which are inte-
grated in a smart phone. The authors explored using
multi-level local pattern (MLP) and local binary pat-
tern (LBP) in feature extraction. For classification,
the authors tried support vector machine (SVM) and
aggregate bootstrapping (bagging). All these models
were evaluated on a 252 gait dataset collected from
42 subjects and yield accuracy up to 77.45% by MLP
and bagging.
2.2 Wavelet Transform on IMU Signals
Researchers have adopted the use of wavelet trans-
form for the analysis of signals. Here, we present
some examples where wavelet transform has proven
to be a very robust approach.
The authors in (Abdu-Aguye and Gomaa, 2019)
used wavelet transform followed by adaptive pooling
to do feature extraction for human activity recogni-
tion based on accelerometer and gyroscope signals.
Their approach was evaluated on seven different ac-
tivity recognition datasets and yielded significant re-
sults.
In (Zhenyu He, 2010), wavelet transform was ap-
plied on 3D acceleration signals then applying an au-
toregressive model on the decomposed signal. The
outcome coefficients were then used as the feature
vector which was fed to a support vector machine
classifier to distinguish between the different human
ICINCO 2020 - 17th International Conference on Informatics in Control, Automation and Robotics
630

activities. This model was tested on four different hu-
man activities and resulted in high accuracy classifica-
tion (95.45%) which clearly shows that the approach
can successfully be used in human activity recogni-
tion based on acceleration signals.
In (Assam and Seidl, 2014), the authors applied
wavelet transform alongside with vector quantization
and Hidden Markov Model (HMM) on sensory data
from Android smart phones. The model aimed to
extract the spectral features of accelerometer sensor
signals by performing multi-resolution wavelet trans-
form and using them for human activity recognition.
The model result was very significant as it reached
classification accuracy up to 96.15% on six human
activities.
From the previous works, it seems evident that
wavelet transform performs well in the analysis of ac-
celerometer and gyroscope signals. However, it was
used only for activity recognition. In our work, we
explore using wavelet transform for gender recogni-
tion and seeing if the signal decomposition still able to
distinguish between male signals and female signals
fixing a particular activity, which is ‘walking’ in the
current work. We also investigate which body part(s)
are the best in uniquely differentiating gender based
on inertial signals. Finally, we investigate the cultural
impact and hypothesize that the effectiveness of iner-
tial signals in gender recognition may be dependent
on the culture from where the subjects come from.
2.3 Random Forests
The work presented in (Mehrang et al., 2018) uses
random forests as the main classifier to recognize hu-
man activity based on triaxial acceleration signals.
The system achieved accuracy of 89.6 ± 3.9% with
a forest of size 64 trees.
The authors in (Feng et al., 2015) designed an
ensemble learning algorithm that integrates many in-
dividual random forest classifiers. Their model was
evaluated on a dataset consisting of 19 different phys-
ical activities and reached accuracy up to 93.44% with
a small training time compared to other classification
methods.
The authors in (Casale et al., 2011) also adopted
the technique of using random forests classifier to rec-
ognize human activities based on acceleration signals.
They obtained a high classification accuracy which
was up to 94%.
From these results, we can conclude that the ran-
dom forests classifier can be efficiently applied on ac-
celerometer and gyroscope signals to perform gender
recognition.
3 EJUST-GINR-1 DATASET
The dataset was collected using six IMU units and
two smart watches. Each IMU unit is a MetaMo-
tionR (MMR) sensor, which is a wearable device
that provides real-time and continuous motion track-
ing (MBIENTLAB, 2018). We record the readings
of the following sensors: Accelerometer, Gyroscope,
Magnetometer and Pressure. The internal compo-
nents of each MetaMotionR sensor is shown in Fig-
ure 1.
Figure 1: The components of each MetaMotionR
unit (MBIENTLAB, 2018).
However, within the scope of this work, we only
use the accelerometer and gyroscope. The MetaMo-
tionR sensor specifications are illustrated in Table 1.
The sensors were placed at six positions: right upper
arm (RUA), left upper arm (LUA), right cube (RC),
left cube (LC), waist and back, along side with two
Apple watches (LH) and (RH) as shown in Figure 2.
The two watches model is series-1, integrated in
them Apple S1 computer which is described as a Sys-
tem in Package (SiP). The SiP includes the two sen-
sors we need: accelerometer and gyroscope. Figure 3
shows the Apple watch series-1 and the sensor axes.
Figure 2: MetaMotionR units and smart watches place-
ments on the human body as indicated by the red spots.
The sensors were synchronized together alongside
with the smart watches to generate gyroscope and
Multi-sensor Gait Analysis for Gender Recognition
631

Table 1: Sensor specification of the MetaMotionR unit.
Description Ranges Resolution
Sample
Rate
Gyroscope
±125,
±250,
±500,
±1000,
±2000
deg/s
16 bit
0.001Hz,
100Hz
stream,
800Hz
log
Accelerometer
±2,
±4,
±8,
±16g
16 bit
0.001Hz,
100Hz
stream,
800Hz
log
Figure 3: Apple watch series-1 used for recording ac-
celerometer and gyroscope signals.
accelerometer readings with frequency equals to 50
Hertz.
The subjects who participated in collecting this
dataset are all volunteer students at our university
(both postgraduate and undergraduate). The total
number of data samples and subjects information are
summarized in Table 2. Gait procedure: Each sub-
ject walked alone on a straight ground for 4 sessions,
each session lasted for 5 minutes, totalling over than
one million sensor readings. The process was stan-
dardized among all subjects. Males and females were
wearing trousers in order not to change the readings
of the sensors placed on the subject’s legs. Partici-
pants were asked to walk naturally in the same way
they walk every day. The dataset is available upon
request.
4 METHODOLOGY
4.1 Feature Extraction
In our domain, we are dealing with timeseries which
are the signals coming from IMU sensors. In or-
der to analyse the timeseries, we would like to know
which frequencies are present in the signal. Fourier
transform is a famous method to do that. However,
Table 2: Total number of samples and subjects information
of EJUST-GINR-1 dataset.
Attribute Value
Total Number of Samples 5292
Number of Females 10
Number of Males 10
Age Range 19-33
Height Range 146cm-187cm
Weight Range 56kg-130kg
Fourier transform doesn’t give any information about
time (it has a high resolution in the frequency-domain
but zero resolution in the time-domain) as explained
in Figure 4. For that reason, scientists proposed the
use of Short-Time Fourier transform. The main prob-
lem with this approach is that we face the same limits
of Fourier Transform known as the uncertainty prin-
ciple. The smaller we make the size of the window
the more we will know about where a frequency has
occurred in the signal, but less about the frequency
value itself (Taspinar, 2018).
Figure 4: An overview of time and frequency resolutions
of various transformations. The size and orientations of the
block gives an indication of the resolution size (Taspinar,
2018).
To solve this problem, Wavelet transforms are
used as they provide high resolution in frequency do-
main and also in time domain. This means we can
know which frequencies are present in a signal and
also at what time these frequencies have occurred.
Unlike Fourier transform, the Wavelet transform
represents a signal as a decomposition of some func-
tions called Wavelets. Each wavelet is at a differ-
ent scale. The difference between wavelets and sine
waves is that wavelets are positioned in time. Mathe-
matically, the Wavelet transform is described by equa-
tion (1):
W (a, b) =
1
p
|a|
Z
∞
−∞
x(t)ψ(
t − b
a
)dt (1)
Here, W (a, b) is the wavelet coefficient, a is the scal-
ICINCO 2020 - 17th International Conference on Informatics in Control, Automation and Robotics
632

ing variable, b is the shifting variable and Ψ(t) is
called the mother wavelet. Practically, these coeffi-
cients are calculated from the correlation between the
signal at a time instance t and a wavelet shifted to
the same time instance. The accuracy of the signal
representation depends on how good we choose the
wavelet. There are various families of wavelets in the
literature such as Haar, Complex Gaussian wavelets,
Morlet, etc.
In this work, we use the Morlet wavelet which is
described by equation (2):
ψ(t) = exp(
−t
2
2
)cos (5t) (2)
Now, the coefficients obtained from Morlet wavelet
transform are considered a good descriptor of the
original signals. We then take these coefficients as
the feature vector of our dataset and feed them to the
classifier.
4.2 Classification
There are many methods in supervised machine learn-
ing that solve classification problems. Some of them
do both feature extraction and classification. How-
ever, we set the feature vector as the outcome from
the wavelet transform, then we consider two different
classifiers: Random forest and Convolutional Neural
Network (CNN).
A. Random Forest. As shown in our literature re-
view, random forests is one of the most popular and
efficient techniques as it is based on ensemble learn-
ing. Ensemble learning means that the model we use
makes predictions based on many different individual
mini-models. Ensemble learning is done in two ways,
either bagging or boosting. Bagging means that the
individual models are trained in parallel and each one
uses a subset of the dataset. A random forest uses
the decision trees as its individual models with bag-
ging. In addition it does random feature selection to
be able to improve the classification accuracy by re-
ducing the prediction variance. In this work, we use
a random forest classifier with its specification shown
in section 5.
B. Convolutional Neural Network. CNN is a very
famous approach to perform classification on multi-
dimensional data. Although CNNs are considered a
deep learning method, we don’t use a deep network
in this work for three reasons. Firstly, we don’t need
the network to learn more complex features as we al-
ready have our feature vector from the Wavelet trans-
form. Secondly, a deep convolutional network will
require much computations and we want to minimize
the computations on the dataset as possible as the cur-
rent research can be later used for online implementa-
tion on wearable devices. Finally, we would need a lot
more data in deep learning to prevent overfitting. For
these reasons, we use a shallow network as a classifier
with its specifications described in section 5.
5 EXPERIMENTAL SETUP
To investigate the effectiveness of our methodology,
we evaluate the performance on two datasets: our own
collected EJUST-GINR-1 dataset and The OU-ISIR
Gait Dataset (Ngo et al., 2014). Both datasets include
accelerometer and gyroscope signals from IMU sen-
sors collected for gait analysis and identification of
human attributes.
5.1 Datasets Considered
5.1.1 Experiments Setup on EJUST-GINR-1
Dataset
Using EJUST-GINR-1 dataset, introduced in sec-
tion 3, we ran some experiments using our model.
Firstly, we use signals coming from each of the eight
sensors individually to know which part(s) of the hu-
man body best distinguishes between males and fe-
males during walking. Then, we use many combina-
tions of sensors to see if this will have an impact on
the predictive performance.
The dataset was split into fixed-size samples.
Each sample corresponds to a 5-second signal with
its label indicating whether the subject is a male or a
female. The sampling rate was 50 Hz so, each sample
had 250 sensor readings with each reading consisting
of six components: Accelerometer-X, Accelerometer-
Y, Accelerometer-Z, Gyroscope-X, Gyroscope-Y, and
Gyroscope-Z.
5.1.2 Experiments Setup on the OU-ISIR Gait
Dataset
The OU-ISIR Gait dataset was collected at Osaka
University to help research in the area of human iden-
tification based on gait analysis (Ngo et al., 2014). We
had the permission to use the dataset in our research
from the dataset administrator with a signed agree-
ment from EJUST University to Osaka University.
The dataset was collected using three IMU sen-
sors and a smart phone, all located around the waist
of the subject. The dataset included three gait styles:
Multi-sensor Gait Analysis for Gender Recognition
633

level walk, up slope and down slope. The dataset in-
cludes readings from IMUs similar to the sensors we
used to collect our dataset. Each unit generates six-
dimensional data: Accelerometer-X, Accelerometer-
Y, Accelerometer-Z, Gyroscope-X, Gyroscope-Y, and
Gyroscope-Z. The dataset was aggregated to differ-
ent versions to satisfy many protocols for different re-
search goals. We used two versions of the dataset.
Firstly, the one with the largest number of subjects
(total 744 subjects). This version has readings from
only the sensor placed on the centre of the waist for
level walk activity with sampling rate equals to 100
Hz, each subject has two signals of level walk. The
total number of samples and subjects information of
level walk dataset version are summarized in Table 3.
The second version has a less number of subjects (to-
tal 495 subjects) and contains two signals of level
walk, one for slope up and one for slope down for
each subject shown in Table 4.
The length of the signals is not included in the
dataset description. However, from the data itself we
can conclude that each signal has a length of only few
seconds.
Table 3: Total number of samples and subjects information
of level walk only OU-ISIR Gait dataset version.
Attribute Value
Total Number of Samples 1488
Total Number of Subjects 744
Age Range 2-78
Table 4: Total number of samples and subjects informa-
tion of level walk, slope up and slope down OU-ISIR Gait
dataset version.
Attribute Value
Total Number of Samples 1980
Total Number of Subjects 495
Age Range 2-78
5.2 Model Specification
Our model for the EJUST-GINR-1 dataset is de-
scribed as follows. First, we feed the fixed length
signals to the feature extractor, which applies Mor-
let wavelet transform. We tried different scales of de-
composition and selected the range of scales that min-
imizes the computations and gives a high accuracy.
The scales range from 1 to 64 gave the best results.
We take all the coefficients obtained from the signal
decomposition and consider them the feature vector
then feed it to the classifier.
In the random forest classifier, we set the number
of decision trees in the forest to 100 trees and use the
Gini index to measure the quality of the split. We
use bootstrap aggregation to randomly select subsets
of the whole dataset and also random subsets of the
features. We run each experiment 10 times and report
the average accuracy.
In the CNN classifier, we use a shallow network
consisting of 5 convolutional layers, each followed by
a max pooling layer, and at the end, one fully con-
nected layer then the final layer that produces the bi-
nary classification output. The activation function at
all layers are ReLU (Rectified Linear Unit) except at
the output layer, we use softmax as activation to pro-
duce the output scores. All the code components are
available on a GitHub repository and available upon
request.
6 RESULTS AND DISCUSSION
In this section, we include the results obtained from
the experimental evaluation of our methodology. We
consider the classification accuracy as the evaluation
criteria to our model.
6.1 Evaluation on the EJUST-GINR-1
Dataset
The results of our approach using wavelet transform
to extract features followed by random forest clas-
sifier evaluated on the EJUST-GINR-1 dataset are
shown in Table 5. In Table 5, we refer to the sen-
sors placed in the left upper arm, right upper arm, left
cube, right cube, left hand and write hand as LUA,
RUA, LC, RC, LH and RH respectively along side
with back and waist sensors without name abbrevia-
tions. The diagonal elements of the table represent the
classification accuracy obtained by testing on the sig-
nals of each sensor individually. The off-diagonal el-
ements represent the classification accuracy obtained
by combining the data of each sensor with the data
of each of the other seven sensors in order. For the
overall performance, all results of individual sensors
lie between 85.96% for the back sensor and 95.85%
for the left cube sensor. In general, we can conclude
from the results that the sensors located at lower part
of the body (right cube, left cube and waist) classify
gender by significantly higher accuracy than the sen-
sors located at the upper part of the body.
Evaluating the model on 12-dimensional data by
combining the accelerometer and gyroscope readings
of each two sensors boosted the performance of indi-
vidual sensors. As illustrated in Table 5, the accuracy
obtained from The left upper arm sensor was 89.06%,
and from right upper arm 86.72% but, when the
two sensors combined together, the accuracy reached
94.26%. The performance also was boosted for many
ICINCO 2020 - 17th International Conference on Informatics in Control, Automation and Robotics
634

Table 5: Accuracy obtained by combinations of sensors.
Sensor LUA RUA LC RC Back Waist LH RH
LUA 89.0578% 94.2647% 94.3103% 94.7267% 89.4118% 92.6782% 92.9979% 93.1703%
RUA 94.2647% 86.7155% 95.9459% 94.6705% 88.8400% 95.0291% 93.8272% 90.6826%
LC 94.3103% 95.9459% 95.8451% 94.7767% 93.8143% 96.7388% 96.5509% 94.4693%
RC 94.7267% 94.6705% 94.7767% 93.4627% 95.0348% 96.0169% 96.5698% 93.6639%
Back 89.4118% 88.8400% 93.8143% 95.0348% 85.9539% 94.8727% 91.0545% 90.8710%
Waist 92.6782% 95.0291% 96.7388% 96.0169% 94.8727% 95.7143% 96.6539% 95.3160%
LH 92.9979% 93.8272% 96.5509% 96.5698% 91.0545% 96.6539% 89.0566% 91.2606%
RH 93.1703% 90.6826% 94.4693% 93.6639% 90.8710% 95.3160% 91.2606% 88.9023%
Table 6: Our approach compared to previous approaches on the OU-ISIR dataset (Garofalo et al., 2019).
Method Classification Accuracy
AutoWeka 2.0 58.25%
HMM 41.75%
TCN 60.31%
TCN + Orientation Independent 67.01%
CNN + AE-GDI 75.77%
Ensembel 64.43%
Wavelet + CNN 70.85%
Wavelet + random forest 68.27%
other combinations. The highest accuracy was ob-
tained by combining the left cube sensor with the
waist sensor with accuracy reaching 96.74%.
6.2 Evaluation on the OU-ISIR Gait
Dataset
We evaluated our methodology on both versions of
the OU-ISIR Gait dataset (Ngo et al., 2014) described
in section 5 using wavelet transform as a feature ex-
tractor then trying both classifiers random forest and
CNN. Both versions of the dataset give relatively
lower results than EJUST-GINR-1 dataset. Our model
evaluated on the first version with the highest num-
ber of subjects (744 subjects) reached 74.73% accu-
racy using CNN classifier specified in section 5, and
69.03% accuracy using the random forest classifier.
In the second version of the dataset, which in-
cludes less number of subjects (495 subjects) but
more gait styles, we obtain accuracy up to 70.85% us-
ing CNN and 68.27% using random forest. We com-
pare our results on the second version of the dataset
to other approaches, all summarized in Table 6. As
shown in Table 6, using wavelet transform as a fea-
ture extractor outperforms most other approaches. We
should also mention that the computation power and
time needed for applying wavelet transform is signif-
icantly less than any deep learning technique.
7 CONCLUSION AND FUTURE
WORK
In this work, we proposed a reliable model for gen-
der recognition based on inertial data of accelerom-
eter and gyroscope signals streamed from wearable
IMU units. We use wavelet transform as our fea-
ture descriptor. We also proposed a new gait dataset
EJUST-GINR-1 collected from smart watches and
IMU sensors placed at eight different parts of the hu-
man body. We investigated which body locations,
in terms of sensors placements, best distinguish be-
tween males and females in walking. We evaluated
our model on two datasets and showed the results
of each dataset. Our approach gives very promising
results which shows that wavelet transform can effi-
ciently be used to extract features for gender recogni-
tion along with potentially diverse set of classifiers.
In the future, we intend to expand our approach to
make age predictions, as both a classification problem
(age level) as well as regression problem (estimate the
exact age), and expand our dataset to include more
age ranges and more activities. We would like to also
investigate which other activities/actions can reliably
differentiate gender using inertial sensors. These ac-
tions can include brushing teeth, sitting, standing, etc.
We may also investigate the use of wavelet transform
to analyse electroencephalogram (EEG) signals to do
gender recognition based on brain signals analysis.
Multi-sensor Gait Analysis for Gender Recognition
635

ACKNOWLEDGMENTS
This work is funded by the Information Technol-
ogy Industry Development Agency (ITIDA), Infor-
mation Technology Academia Collaboration (ITAC)
Program, Egypt – Grant Number (PRP2019.R26.1 - A
Robust Wearable Activity Recognition System based
on IMU Signals).
We also would like to thank all the students who
participated in collecting the dataset.
REFERENCES
Abdu-Aguye, M. G. and Gomaa, W. (2019). Competi-
tive feature extraction for activity recognition based
on wavelet transforms and adaptive pooling. In 2019
International Joint Conference on Neural Networks
(IJCNN), pages 1–8.
Assam, R. and Seidl, T. (2014). Activity recognition from
sensors using dyadic wavelets and hidden markov
model. In 2014 IEEE 10th International Conference
on Wireless and Mobile Computing, Networking and
Communications (WiMob), pages 442–448.
Casale, P., Pujol, O., and Radeva, P. (2011). Human activity
recognition from accelerometer data using a wearable
device. In Iberian Conference on Pattern Recognition
and Image Analysis, pages 289–296. Springer.
Feng, Z., Mo, L., and Li, M. (2015). A random forest-based
ensemble method for activity recognition. In 2015
37th Annual International Conference of the IEEE En-
gineering in Medicine and Biology Society (EMBC),
pages 5074–5077. IEEE.
Garofalo, G., Argones R
´
ua, E., Preuveneers, D., Joosen, W.,
et al. (2019). A systematic comparison of age and gen-
der prediction on imu sensor-based gait traces. Sen-
sors, 19(13):2945.
Jain, A. and Kanhangad, V. (2016). Investigating gen-
der recognition in smartphones using accelerometer
and gyroscope sensor readings. In 2016 International
Conference on Computational Techniques in Infor-
mation and Communication Technologies (ICCTICT),
pages 597–602.
Lu, J., Wang, G., and Moulin, P. (2014). Human identity
and gender recognition from gait sequences with arbi-
trary walking directions. IEEE Transactions on Infor-
mation Forensics and Security, 9(1):51–61.
MBIENTLAB (2018). https://mbientlab.com.
Mehrang, S., Pietil
¨
a, J., and Korhonen, I. (2018). An activ-
ity recognition framework deploying the random for-
est classifier and a single optical heart rate monitor-
ing and triaxial accelerometer wrist-band. Sensors,
18(2):613.
Ngo, T. T., Ahad, M. A. R., Antar, A. D., Ahmed, M., Mura-
matsu, D., Makihara, Y., Yagi, Y., Inoue, S., Hossain,
T., and Hattori, Y. (2019). Ou-isir wearable sensor-
based gait challenge: Age and gender. In Proceedings
of the 12th IAPR International Conference on Biomet-
rics, ICB.
Ngo, T. T., Makihara, Y., Nagahara, H., Mukaigawa, Y.,
and Yagi, Y. (2014). The largest inertial sensor-
based gait database and performance evaluation of
gait-based personal authentication. Pattern Recogni-
tion, 47(1):228–237.
Rosli, N. A. I. M., Rahman, M. A. A., Balakrishnan, M.,
Komeda, T., Mazlan, S. A., and Zamzuri, H. (2017).
Improved gender recognition during stepping activity
for rehab application using the combinatorial fusion
approach of emg and hrv. Applied Sciences, 7(4):348.
Shepstone, S. E., Tan, Z.-H., and Jensen, S. H. (2013). De-
mographic recommendation by means of group profile
elicitation using speaker age and gender recognition.
In INTERSPEECH, pages 2827–2831.
Taspinar, A. (2018). A guide for using the
wavelet transform in machine learning.
http://ataspinar.com/2018/12/21/a-guide-for-using-
the-wavelet-transform-in-machine-learning.
Yuchimiuk, J. (2007). System and method for gender iden-
tification in a speech application environment. US
Patent App. 10/186,049.
Zhang, K., Gao, C., Guo, L., Sun, M., Yuan, X., Han, T. X.,
Zhao, Z., and Li, B. (2017). Age group and gender
estimation in the wild with deep ror architecture. IEEE
Access, 5:22492–22503.
Zhenyu He (2010). Activity recognition from accelerome-
ter signals based on wavelet-ar model. In 2010 IEEE
International Conference on Progress in Informatics
and Computing, volume 1, pages 499–502.
ICINCO 2020 - 17th International Conference on Informatics in Control, Automation and Robotics
636